Authors

Abstract

To address the limitations of computational resources and the stringent real-time stability requirements in post-harvest online grading of fresh tea leaves, this study proposes a lightweight object detection method oriented toward efficient deployment. The YOLOv11n network was selected as the baseline model. A StarNet backbone was introduced to enhance nonlinear inter-channel interactions while reducing the complexity of feature extraction. In addition, a DySample dynamic upsampling module was employed to predict content-adaptive sampling locations, thereby improving multi-scale feature reconstruction under lightweight constraints. Furthermore, a wavelet-based pooling structure was designed to perform structure-aware frequency-domain decomposition, preserving critical edge and texture information while reducing redundant computation. The Inner-MPDIoU loss function was also adopted to calculate overlap consistency within a compact core region, thereby improving bounding-box localization stability for fine-grained structures. Based on these integrated improvements, the lightweight YOLOv11-SDWI detection model was developed. Experimental results demonstrated that the proposed model achieved an accuracy of 89.5%, with only 4.9 GFLOPs and 1.89 M parameters. To further verify its engineering applicability, the algorithm was encapsulated into a complete visual software system specifically designed for fresh tea leaf sorting. The developed system provides a functional interface for real-time detection, effectively bridging theoretical algorithm design and practical agricultural applications, and establishing a solid software foundation for future deployment on physical sorting equipment.

Abstract in English

为了解决采后鲜茶在线分级中计算资源有限以及对高实时稳定性的严格要求，本文研究了一种面向高效部署的轻量化目标检测方法。以YOLOv11n网络为基线模型。引入StarNet主干网络以增强非线性通道间交互并降低特征提取复杂度。采用DySample动态上采样模块来预测内容自适应的采样位置，从而在轻量化约束下改善多尺度特征重建。进一步设计了基于小波的池化结构来执行结构感知的频域分解，在保留关键边缘和纹理信息的同时减少冗余计算。此外采用Inner MPDIoU损失在紧凑的核心区域上计算重叠一致性，增强了细粒度结构的边界框定位稳定性。基于这些系统性改进，构建了YOLOv11 SDWI轻量化检测模型。实验结果表明，所提模型实现了89.5%的精度，同时仅需4.9 G浮点运算数和1.89 M参数量。为了验证其工程应用潜力，该算法被封装成专为鲜茶分拣量身定制的完整视觉软件系统。这种系统级实现为实时检测提供了功能性界面，成功弥合了理论算法设计与实际农业应用之间的差距，并为未来在物理分拣设备上的部署奠定了坚实的软件基础。